You are on page 1of 11

Applied Soft Computing 34 (2015) 827837

Contents lists available at ScienceDirect

Applied Soft Computing


journal homepage: www.elsevier.com/locate/asoc

A tabu search based hybrid evolutionary algorithm


for the max-cut problem
Qinghua Wu a , Yang Wang b, , Zhipeng L c
a
School of Management, Huazhong University of Science and Technology, 430074 Wuhan, China
b
School of Management, Northwestern Polytechnical University, 127 Youyi West Road, 710072 Xian, China
c
School of Computer Science and Technology, Huazhong University of Science and Technology, 430074 Wuhan, China

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents a tabu search based hybrid evolutionary algorithm (TSHEA) for solving the max-cut
Received 2 December 2014 problem. The proposed algorithm integrates a distance-and-quality based solution combination operator
Received in revised form 1 April 2015 and a tabu search procedure based on neighborhood combination of one-ip and constrained exchange
Accepted 11 April 2015
moves. Comparisons with leading reference algorithms from the literature disclose that the proposed
Available online 14 May 2015
algorithm discovers new best solutions for 15 out of 91 instances, while matching the best known solu-
tions on all but 4 instances. Analysis indicates that the neighborhood combination and the solution
Keywords:
combination operator play key roles to the effectiveness of the proposed algorithm.
Max-cut
Metaheuristics 2015 Elsevier B.V. All rights reserved.
Hybrid evolutionary algorithm
Tabu search

1. Introduction in [7,13,22]. Due to the computational complexity of solving large scale problem
instances, various heuristic algorithms are accordingly proposed. Geomans and
Given an undirected graph G = (V, E), each edge e(i, j) E being associated with Williamson proposed a semidenite relaxation based randomized heuristic and
a weight wij , the max-cut problem is to partition the vertex set V into two disjoint established 0.878 performance guarantee [9]. Burger et al. proposed a rank-two
subsets V0 and V1 , without requiring that |V0 | = |V1 |, such that the total weights of all relaxation heuristic that does not increase the number of variables when approxi-
the cutting edges are maximized. A cutting edge is such an edge that connects two mating large-scale problems in despite of a non-convex relaxation [6]. Festa et al.
vertices with one in V0 and the other in V1 . The objective function of the max-cut investigated several heuristics derived from greedy randomized adaptive search
problem can be represented as follows: procedure, variable neighborhood search and path relinking [8]. Palubeckis and
 Krivickiene designed two multistart tabu search implementations, one of which
f (V0 , V1 ) = max wij (1) produces initial solutions by xing variables and the other operates in a random
way [19]. Krishnan and Mitchell proposed a semidenite programming based poly-
iV0 ,jV1
hedral cut and price approach, in which the pricing phase is solved using the interior
Alternatively, let the variable xi represent the vertex i. Let xi receive the value point cutting plane algorithm and the cutting phase is based on the polyhedral the-
0 if the vertex i belongs to V0 and receive the value 1 otherwise. Then the max-cut ory [15]. Mart et al. proposed a scatter search algorithm which is characterized of
problem can be formulated with the following binary quadratic function: constructing a reference set by the way of solving the maximum diversity prob-
lem, conducting an ejection chain move in the local search, and applying multiple

n

n types of combination methods [16]. Arriz and Olivo implemented tabu search and
f (x) = max wij (1 xi )xj (2) simulated annealing heuristics, both of which use the linearly selected move, ran-
dom move and their combination [2]. Kochenberger et al. reformulated the max-cut
i=1 j=1,j =
/ i
problem into the binary quadratic programming model (BQP) and solved the refor-
where n denotes the number of vertices in the graph G. mulated problem with a diversication driven tabu search algorithm [14]. Wang
The solution of the max-cut problem consists in determining a binary value (0/1) et al. also reformulated the max-cut problem as the BQP problem and proposed
for each variable in a n-dimensional vector. two path relinking algorithms, where one employs a greedy strategy to generate
The max-cut problem is one of Karps 21 NP-complete problems and has a path of intermediate solutions connecting two solutions and the other operates
important applications in circuit layout design and statistical physics [3]. The past in a random way [27]. Shylo et al. developed global equilibrium search algorithms
decades have witnessed many solution procedures for tackling the max-cut prob- consisting of initial solution construction dependent on accumulated information
lem reported in the literature. Several examples of exact algorithms can be found from a set of good solutions and tabu search for solution improvement [23,24]. Ben-
lic and Hao proposed a breakout local search algorithm which alternates between a
steepest descent phase to discover local optima and an adaptive perturbation phase
to locate search into new promising area [5]. Wu and Hao developed a memetic
Corresponding author. Tel.: +86 13891922885. algorithm integrating a multi-parent crossover operator which tries to preserve the
E-mail addresses: qinghuawu1005@gmail.com (Q. Wu), yangw@nwpu.edu.cn largest common vertex groupings with respect to parent solutions and an iterated
(Y. Wang), zhipeng.lv@hust.edu.cn (Z. L). tabu search improvement procedure [28]. Among the reported heuristic algorithms,

http://dx.doi.org/10.1016/j.asoc.2015.04.033
1568-4946/ 2015 Elsevier B.V. All rights reserved.
828 Q. Wu et al. / Applied Soft Computing 34 (2015) 827837

breakout local search, global equilibrium search and path relinking are the best 2.2. Population initialization and updating
heuristics for solving challenging max-cut problems.
In this paper, we present a hybrid evolutionary algorithm featuring a combined
The initial population contains |P| solutions, each of which is
neighborhood based tabu search and a solution combination operator being an inno-
vative variant of the traditional crossover. The tabu search component repeatedly generated as follows. First, a solution x is randomly created, where
alternates between a one-ip move guided tabu search phase and a constrained each variable xi of x receives the value 0 or 1 with an equal proba-
exchange move guided tabu search phase. The proposed solution combination oper- bility of 0.5. Notice that any x is a valid solution. Then this solution
ator is simple in design, which inherits maximally common solution components x is passed to the tabu search procedure for an improved solution
from parent solutions and completes the solution construction using a distance-
and-quality evaluation mechanism.
x . The resulting improved solution is added to the population if it
The proposed tabu search based hybrid evolutionary algorithm (TSHEA) is dis- is not identical to any solution currently in the population. The two
tinct from the previously proposed memetic algorithm (MA) [28] in the following solutions, say x1 and x2 , are determined to be identical if the Ham-
aspects: (1) TSHEA uses neighborhood combination in its tabu search procedure, ming distance d(x1 , x2 ) = 0 or d(x1 , 1 x2 ) = 0 (due to the symmetry
which enhances the neighborhood exploitation in MA that employs a single neigh-
caused by the binary vector representation for the solution of the
borhood; (2) TSHEA uses a solution combination operator similar to the traditional
uniform crossover by reference to two parent solutions while MA uses a group-based max-cut problem). This procedure is repeated until the population
crossover operator by reference to multiple parent solutions; (3) TSHEA is evaluated is lled up with 3 |P| solutions from which we nally retain the
on larger benchmark instances and is capable of achieving better performance than |P| best ones with the largest objective values to form the initial
MA, in particular leading to the improvement of solution quality on several hard population. This procedure allows us to obtain an initial popula-
instances.
tion of relatively high quality. Notice that for some small (or easy)
Experimental comparisons with leading reference algorithms in the literature
reveal that the proposed algorithm is capable of discovering improved solutions for instances, our algorithm can even reach the optimal solutions (or
15 out of 91 instances and matching all but 4 instances. Furthermore, additional solutions of previously best known objective values) during the
analysis discloses the merit of using neighborhood combination and the solution population initialization phase due to the high efciency of the tabu
combination operator for the proposed TSHEA algorithm.
search procedure.
The rest of the paper is organized as follows. Section 2 describes the proposed
algorithm. Section 3 presents experimental results and comparisons with state-of- Once a population is complete, the combination operator that
the-art algorithms in the literature. Section 4 analyzes several essential components produces an offspring solution from selected parent solutions
in the algorithm. Concluding remarks are given in Section 5. begins to work. Each time a new offspring solution is produced
and improved by the tabu search procedure, a population updating
procedure follows. As in the simple version of the scatter search
2. Tabu search based hybrid evolutionary algorithm template [10], the improved offspring solution that is permitted
to be put into the population should satisfy a population updating
Hybrid evolutionary algorithms are known to be particularly rule. This rule only permits a solution to join in the population if
effective for solving many hard combinatorial optimization prob- it does not duplicate any solution and gets better tness (objective
lems [12,18]. Similar to traditional evolutionary algorithms, hybrid function value) than that of at least one solution in the population.
evolutionary algorithms use a crossover operator over parent solu- A qualied solution goes into the population in place of the solution
tions to produce offspring solutions. However, the inclusion of a with the worst solution quality.
local search procedure, such as tabu search [10], for extensively
exploiting the search area around the offspring solutions, is of great
importance to hybrid evolutionary algorithms superiority over tra- 2.3. Parent selection
ditional evolutionary algorithms [4,20].
To maintain an appropriate diversity of the population and to
prevent the search process from premature convergence, we use a
2.1. Main framework parent selection rule to favor the selection of distanced parents for
mating. Our parent selection rule is based on the Hamming distance
d(xi , xj ) between two parent solutions xi and xj , instead of their
Algorithm 1. Outline of the tabu search based hybrid evolutionary tness. The Hamming distance d(xi , xj ) is dened as the number of
algorithm variables that receive different values in solutions xi and xj . In the
1. P = {x1 , . . ., x|P| } Init Pop()
case that the number of such variables is larger than n/2 where n is
2. Record the best solution x* with its cut f(x* )
3. repeat the number of variables, we change all solution components of xi to
4. xp1 , xp2 Select Parents(P) their corresponding complementary values. This is due to the fact
5. xo Combination Operator(xp1 , xp2 ) that we always use the shortest Hamming distance between the
6. xo Tabu Search(xo )
two solutions which are represented with either the binary vector
7. P = {x1 , . . ., x|P| } Population Update(x1 , . . ., x|P| , xc )
8. if f(xo ) > f(x* ) then itself or its complementary vector. The solutions xi and xj chosen as
parents need to satisfy the criterion that their Hamming distance
9. x* = xo , f(x* ) = f(xo )
|P| |P|
10. end if d(xi , xj ) > d, with d = (2/|P|(|P| 1)) i=1 j=i+1 d(xi , xj ) being the
11. until the stopping criterion is satised
average distance between pairs of solutions in the population.
The general scheme of our tabu search based hybrid evolution- Both solution quality and population diversity are essential to
ary algorithm for tackling the max-cut problem is described in the performance of a hybrid evolutionary algorithm. Our distance
Algorithm 1. Starting from an initial population P, TSHEA selects based parent selection rule is used to ensure the diversity of the
two solutions from P as parents to produce an offspring solution, population while our quality based replacement strategy is able to
whereupon a tabu search procedure with a randomized pertur- introduce more high quality solutions into the population. In this
bation mechanism for diversication is launched to optimize the way, we could maintain a healthy population with good diversity
newly generated offspring solution. Subsequently, a population and high quality during the search process.
updating rule decides whether such an improved offspring solution
should be inserted into the population and which existing individ- 2.4. Combination operator
ual in the population should be replaced if yes. The above procedure
repeats until the prescribed stopping criterion is satised. In the fol- Given two selected parents xp1 and xp2 , our tabu search based
lowing sections, we describe the main components of the proposed hybrid evolutionary algorithm uses a dedicated solution combina-
algorithm. tion operator to generate a promising offspring solution xo . The
Q. Wu et al. / Applied Soft Computing 34 (2015) 827837 829

proposed combination operator consists of the following two steps. 12. Identify the variable xi NC with the highest contribution value
The rst step is to create a partial solution by conserving variables when assigned with the same value as xic
13. xio = xio {xic }, NC = NC \ {xi }
that receive the same values in the two selected parent solutions.
14. Update all gt0 and gt1 values for t NC according to Eqs. (3) and (4)
In order to obtain a feasible solution, the second step is to stepwise 15. end for
assign the unassigned variables to complete this partial solution,
making use of a diversication-guided strategy that takes into 2.5. Tabu search
account both the solution quality of the offspring and its distance
to its parent solutions. Tabu search is a powerful metaheuristic for solving many combi-
Relative to a given incomplete solution x = (x1 , x2 , . . ., xp ) and a natorial optimization problems [10] and has proved to be effective
variable xi where xi /x, we dene the contribution value of assign- for tackling the max-cut problem [23,24,27,28]. In contrast to previ-
ing xi = 1 as VC 1i (x) = w . By the same token, we dene
x x,x =0 ij ous algorithms that explore a single neighborhood, our tabu search
j j

the contribution value of assigning xi = 0 as VC 0i (x) = xj x,xj =1
wij . employs a combined neighborhood based on the one-ip and
exchange moves to enhance neighborhood exploitation. The com-
To quickly evaluate the contribution value for each variable xi
bined neighborhood was rst proposed in variable neighborhood
not in x, we maintain a vector g0 with each entry gi0 being initialized
search, which switches among different neighborhoods to escape
as VC 0i (x) and a vector g1 with each entry gi1 being initialized as from local optima [17]. Our tabu search procedure repeatedly alter-
VC 1i (x). nates between two complementary neighborhood searches, one
Let Cand denote the set of variables not in x. If variable xk Cand dened by the well known one-ip move and the other dened by
is picked with the assigned value 1 (chosen according to its contri- an innovative constrained exchange move. The following describes
bution value) to join in x, which is then readily removed from Cand, in detail the one-ip move, the exchange move and how to combine
then gi0 for each variable xi in Cand should be updated as follows: them to build the tabu search.

gi0 = gi0 + wik , xi Cand (3) 2.5.1. One-ip move


Given a solution x = (x1 , x2 , . . ., xn ), the one-ip move consists
of changing the value of a solution component xi into its comple-
If variable xk Cand is picked to join in x with the assigned value
mentary value 1 xi (i.e., moving a vertex from its current subset
0, then gi1 should be updated as follows:
into the other subset). After performing the one-ip move w.r.t. the
solution component xi , we get the following solution:
gi1 = gi1 + wik , xi Cand (4)
x = x Flip(i) = (x1 , . . ., xi1 , 1 xi , xi+1 , . . ., xn ) (5)

The pseudo-code of the solution combination operator is pre- The neighborhood N1 of the solution x consists of all such solu-
sented in Algorithm 2, which mainly consists of two steps. The tions that are obtained by performing the one-ip move for each
rst step is to construct a partial solution xo by inheriting maxi- solution component xi , which contains a total of n neighboring
mally common solution components from the two selected parent solutions.
solutions xp1 and xp2 . Lines 56 in Algorithm 2 are used because The move gain represents the difference in the objective func-
of the symmetry of the binary vector representation. The second tion value between the resulting neighboring solution after a move
step gradually enlarges the partial solution xo by a sequence of is performed and the current solution. To quickly calculate move
construction steps until a complete solution is acquired. gains, we use the similar stream-lined technique as in the uncon-
Each constructive step enlarges the partial solution xo by joining strained binary quadratic optimization problem [11]. Specically,
a variable together with its assigned value from the parent solution given a solution x = (x1 , x2 , . . ., xn ), we maintain a vector  with
xc , for c = p1 or p2. More precisely, let NC denote the set of variables elements i recording the move gain of ipping the solution com-
not in xo . Considering rst the parent solution xp1 , we identify the ponent xi . Then the vector  can be initialized as follows:
variable xi NC with the highest contribution value when assigned 
p1 p1 jV0 wij jV1 wij (xi = 0)
to the same value as in xi and include xi into xo . Afterward, we i = (6)
identify the variable xi NC using the same rule and include xi into
p2 jV1 wij jV0 wij (xi = 1)
xo . Each variable joined into xo is immediately removed from NC.
After performing the one-ip move w.r.t. the solution compo-
The parent solutions xp1 and xp2 are alternatively chosen until NC
nent xi , the  is updated as follows:
becomes empty. One observes that the offspring xo generated by

this combination operator has essentially the same Hamming dis- k (k = i)
tance |NC|/2 to its two parents, thus keeping good diversication.

Meanwhile, the objective contribution based construction proce- k = k 2wik (k =
/ i, xk = xi ) (7)


dure enables the offspring xo hold good solution quality. k + 2wik (k =
/ i, xk = 1 xi )

Algorithm 2. Outline of the combination operator With this additional memory, the complexity to identify the best
1. Input: Two parent solutions xp1 and xp2 one-ip move reduces to O(n) from O(n2 ). Notice that we only need
2. Output: An offspring solution xo to maintain one such  vector to record the move gains for all solu-
3. Let V denote the set of all the variables and n denote the size of V
tions during the search process. After initialization, we can measure
4. Identify the set of maximally common variables C between xp1 and xp2 ,
i.e. C = xp1 xp2 the objective difference between the neighboring solution and the
5. if |C| < n/2 then current solution by only updating this vector.
6. C = xp1 (1 xp2 )
7. end if
p1 2.5.2. Exchange move
8. Create a partial solution by setting xio = xi if xi C
9. Identify the set of noncommon variables NC = V \ C When the tabu search based on the one-ip move neighbor-
10. for k = 1, . . ., |NC|do hood is detected being trapped in a local optimum, our algorithm
11. Select parent xc to be considered: if k is odd, then c = p2, otherwise switches to the exchange move based tabu search for further
c = p1 improvement. The idea of using multiple neighborhoods comes
830 Q. Wu et al. / Applied Soft Computing 34 (2015) 827837

from the fact that a local minimum within a neighborhood is not 19. /* Make improvement with the exchange move based tabu search phase
necessarily a local optimum within another neighborhood and */
20. repeat
an algorithm that explores multiple neighborhoods is expected 21. Identify the best non-tabu exchange move or the best tabu move that
to have more chances to nd better solutions [17]. Although satises the aspiration rule, say this move corresponds to exchange xi and xj
the exchange move has been widely exploited for solving the 22. xi = 1 xi , xj = 1 xj
maximum clique problem [30], traveling salesman problem [26], 23. f (x) = f (x) + i + j + 2wij
24. Update  according to Eq. (10)
quadratic assignment problem [25], etc., this is the rst study of
25. Update Tabu List by setting TLi = Iter + tl, TLj = Iter + tl
using the exchange move to the max-cut problem. 26. iff(x) > f(x* )then
The exchange move consists of exchanging two solution compo- 27. x* = x, f(x* ) = f(x), NonImpN2 = 0
nents xi and xj , subject to requiring that xi and xj receive different 28. else
assigned values (i.e., moving the vertex i from its current subset, 29. NonImpN2 = NonImpN2 + 1
30. else if
say V0 , to the other subset V1 and then moving the vertex j from V1
31. Iter = Iter + 2
to V0 ). After performing the exchange move w.r.t. xi and xj , we get 32. until NonImpN2 <
the following solution: 33. end while

Starting from a given initial solution, the proposed tabu search


x = x Exg(i, j) = (x1 , . . ., xi1 , xj , xi+1 , . . ., xj1 , xi , xj+1 , . . ., xn ) algorithm rst conducts the one-ip move based tabu search phase,
in which each iteration performs either the best one-ip move that
(8)
is not tabu or a tabu move that satises the aspiration rule. A move
is set to be tabu if it was performed during the past tl iterations.
The neighborhood N2 of the solution x consists of all such solu- Although this tabu rule is able to prevent the search from going
tions that are obtained by performing the exchange move for each back to previously visited solutions, it may neglect some unvisited
pair of solution components with different assigned values, which high quality solutions. Hence, an aspiration rule is accompanied
contains a total of |V1 | |V0 | neighboring solutions. that permits a move to be performed if performing it leads to a
The move gain ij of performing the exchange move (xi , xj ) is solution with a better objective function value than the best found
calculated as follows: so far.
As long as the best solution cannot be improved within con-
ij = i + j + 2wij (9)
secutive iterations by using the one-ip move, the exchange move
After performing this exchange move, the  value is updated in based tabu search phase is followed to further enhance the solu-
the following way: tion quality. Since the identication of the best exchange move is of
high computational complexity, we carry out the following heuris-
k 2wij (k = i or k = j)
tic mechanism to reduce the computational efforts. Among all the
k = k 2wik + 2wjk (k =
/ i, k =
/ j, xk = xi ) (10) variables with assigned value 1, we pick the ones that are not tabu

and get the best one-ip move gain i (see Section 2.5.1). Like-
k + 2wik 2wjk (k =
/ i, k =
/ j, xk = xj ) wise, among all the variables with assigned value 0, we pick the
ones that are not tabu and get the best j . The exchange move is
2.5.3. Local search procedure only operated on the variables xi and xj . Similar to the tabu rule
Essentially, our tabu search algorithm repeatedly alternates used in the one-ip based tabu search phase, a variable is set to be
between a one-ip move based tabu search phase and an exchange tabu if it is involved in a performed exchange move during the past
move based tabu search phase until a maximum allowed number tl iterations. An exchange move is said to be tabu if either variable
of iterations () is reached. One iteration is counted when a solu- involved in this move is tabu.
tion component xi is ipped into its complementary value 1 xi . In our tabu search, each time a variable xi is ipped, it is for-
Notice that an exchange move can be trivially decomposed into two bidden to be ipped again for the next tl iterations, where tl is a
sequential one-ip moves (i.e., ipping xi to 1 xi and ipping xj to parameter called tabu tenure. For the current study, we selected
1 xj ), which is thus counted as two iterations. The pseudo-code of to set tl = rand[3, . . ., 0.15|V|], where |V| is number of variables and
the tabu search is shown in Algorithm 3. rand[3, . . ., 0.15|V|] takes a random integer from 3 to 0.15|V|. This
Algorithm 3. Outline of the tabu search algorithm randomized function is called each time one move is performed.
1. Input: a given solution x with its solution value f(x) While the exchange move based tabu search phase is not able to
2. Output: the local optimal solution x* found by tabu search with its solution reach a better solution within consecutive iterations, the search
value f(x* ) switches back to the one-ip move based tabu search phase for
3. TL is an n-dimensional vector that maintains the tabu list,  is an
further improvement. This process is repeated until the specied
n-dimensional vector that maintains the move gain of performing each
one-ip move, Iter = 0 maximum number of iterations is reached.
4. while Iter <  ( is a parameter) do
5. Initialize  according to Eq. (6), TLi = 0 for all i = 1 to n, NonImpN1 = 0,
NonImpN2 = 0 3. Computational results and comparisons
6. /* Make improvement with the one-ip move based tabu search phase*/
7. repeat 3.1. Test instances
8. Identify the best non tabu one-ip move or the best one-ip move that is
tabu but satises the aspiration rule, say this move corresponds to ipping xi
The proposed tabu search based evolutionary algorithm
9. xi = 1 xi , f(x) = f(x) + i
10. Update  according to Eq. (7) (TSHEA) is evaluated on two sets of benchmark instances.
11. Update Tabu List by setting TLi = Iter + tl (tl is a parameter) The rst benchmark consists of 71 G-set instances with nodes
12. if f(x) > f(x* ) then ranging from 800 to 20,000, which can be downloaded at
13. x* = x, f(x* ) = f(x), NonImpN1 = 0
http://www.stanford.edu/yyye/yyye/Gset/. These instances are
14. else
15. NonImpN1 = NonImpN1 + 1 generated by a machine-independent graph generator, compris-
16. else if ing of toroidal, planar and random weighted graphs. Many authors
17. Iter = Iter + 1 including [6,8,16,19,23,14,27] employ these instances to test their
18. until NonImpN1 < ( is a parameter) algorithms for the max-cut problem. The second benchmark
Q. Wu et al. / Applied Soft Computing 34 (2015) 827837 831

Table 1 instances where TSHEA fails to reach the best solution values over
Settings of important parameters.
all the runs. However, since all the reference algorithms used for
Parameters Section Description Values comparisons (such as BLS [5], GES [23,24] and PR2 [27]) have 20
|P| 2.2 Population size 10 runs for each instance, we also report the best found solution over
 2.5 Number of TS 1,000,000 20 runs in order to make a fair comparison. In addition, TSHEA gets
iterations after 20/20 hits for 62 out of 91 instances and for those failed instances
recombination the average solution values are quite close to the best ones. Finally,
tl 2.5 Tabu tenure rand[3, . . ., 0.15|V|]
TSHEA achieves good performance with reasonable computational
2.5 Number of consecutive 1000
non-improvement effort. For instances with no more than 3000 variables, the compu-
iterations before tational consumption ranges from 0.08 to 1889.23 s. For instances
switching to another with 30005000 variables, the computational consumption ranges
neighborhood
from 1903.68 to 7125.31 s. In sum, these results disclose that the
proposed TSHEA algorithm performs quite well in terms of both the
solution quality and the computational time.

consists of 20 cubic lattices instances with 1000 or 2744 nodes,


available at http://www.optsicom.es/maxcut/ 3.4. Comparisons with the best performing algorithms

3.2. Parameter settings In order to further show the effectiveness of our TSHEA algo-
rithm, we compare TSHEA with the best performing heuristics in
The parameter settings of TSHEA used in our experiments are the literature, including Breakout Local Search (BLS) [5], Global
given in Table 1. These parameter values were determined by per- Equilibrium Search (GES) [23,24] and a Binary Quadratic Program-
forming a preliminary experiment on a selection of 11 problem ming based Path Relinking (PR2) [27]. Although the Scatter Search
instances (G32 , G33 , G35 , G40 , G41 , G42 , G55 , G57 , G62 , G65 , G66 ) from [16] and CirCut [6] are popular algorithms, they are recently out-
the G-set benchmark instances. The selected instances are typically performed by the reference algorithms we use for comparisons
hard for existing max-cut algorithms and are thus appropriate for [5,23,24,27].
the purpose of comparisons. The chosen 11 instances are used both Tables 4 and 5 show the best and average solution values
for the parameter tuning and for the analysis of key components in obtained by TSHEA, BLS, GLS and PR2 algorithms, respectively.
Section 4. We do not include comparisons with BLS on the cubic lattices
The parameters are tuned as follows. For the parameter |P| instances because BLS do not report results on this set of bench-
required by our memetic algorithm, we simply set it to 10 as many mark instances. Considering that a completely fair comparison of
memetic algorithms do in the literature [12,28,29]. For the three the computing time is quite difcult due to implementation dif-
parameters used in the TS procedure (, tl and ), we tested differ- ferences and different software and hardware platforms used, it is
ent values for each of them, while xing the rest with the values thus preferable to compare algorithms based on the solution qual-
shown in Table 1. The parameter settings are further justied by ity. For indicative purpose, TSHEA is highly competitive with the
the parameter sensitivity analysis in Section 4.3. other algorithms in terms of computational efciency. The results
Our TSHEA algorithm is programmed in C++ and compiled using unavailable in Table 4 are marked as and the best results in each
GNU g++ on a Xeon E5440 with 2.83 GHz CPU and 8 GB RAM. Given row are marked as bold.
the stochastic nature of our algorithm, each instance is solved for 20 From Table 4, we nd that for the rst 54 instances, TSHEA is
independent runs with different random seeds. Each run is stopped able to nd better results for 6, 3 and 17 instances than BLS, GES
when the processing time reaches the prescribed time limit, which and PR2 respectively. For the 17 large G-set instances, TSHEA is
is set to be 30 min for graphs with |V| < 5000 and 120 min for graphs able to nd better results for 12 instances while it fails to match
with |V| 5000. Notice that this time limit is quite comparable with the results of BLS for only 4 instances. Furthermore, TSHEA is quite
what the reference algorithms use [5,16,27]. robust to reach the best solutions for the former 54 small or medium
instances, which is much better than BLS and PR2. For the large
3.3. Computational statistics instances, it is still hard for TSHEA as well as other reference algo-
rithms to have a robust performance. In addition, in order to see
Tables 2 and 3 present the computational statistics of the whether there is any statistical difference between the population
proposed TSHEA algorithm on the G-set and the cubic lattices of the 20 results achieved by our TSHEA and the mean reported by
benchmark instances, respectively. In each table, columns 1 to 3 the reference algorithms, we conduct the one-sample t-test with
give the instance names (Instance), number of vertices (|V|) and a signicance factor of 0.05 for each problem instance. Since this
the best known results (fprev ) previously reported in the literature statistical test requires the standard deviation on a set of results is
[5,23,24,27,28]. The columns under the heading TSHEA report the not 0, we add a random value taken from [ 0.0001, 0.0001] to such
best solution values (fbest ), the average solution values (favg ) along results that are the same during all the 20 runs for one instance. The
with the standard variance shown in parenthesis (), the number statistical results reveal that TSHEA is statistically better than BLS
of hits to reach the fbest over 20 runs (hits), the total number of iter- and PR2 for most instances and signicantly outperforms GES for 8
ations Iters and the average time (time) to reach fbest (in seconds). instances with p value < 0.05.
Results marked in bold indicate that TSHEA gets better results than From Table 5, we observe that TSHEA is able to reach better
the best known results fprev . results than GES and PR2 for 2 and 10 out of 20 cubic lattices
From Tables 2 and 3, we rst observe that TSHEA is capable of instances, respectively. In addition, the average values of TSHEA
improving the best known results for a total of 15 instances, among are competitive with GES and much better than PR2. Similar to the
which are 13 G-set instances and 2 cubic lattices instances. More- Table 4, we conducted the one-sample t-test for each instance and
over, TSHEA matches the best known results for all instances except the results reveal that TSHEA performs signicantly better than PR2
4 instances G62, G72, G77, G81. Given the random nature of our and GES for 17 and 2 instances, respectively. In conclusion, com-
algorithm, its possible to obtain better results in terms of the best parisons with state-of-the-art algorithms on a total of 91 instances
solution values if each instance has more runs, in particular to large demonstrate the effectiveness of our proposed TSHEA algorithm.
832 Q. Wu et al. / Applied Soft Computing 34 (2015) 827837

Table 2
Computational results of TSHEA on the G-set instances.

Instance |V| fprev TSHEA

fbest favg () Hits Iter Time

G1 800 11, 624 11, 624 11, 624 20/20 237,516 1.10
G2 800 11, 620 11, 620 11, 620 20/20 1,380,012 4.70
G3 800 11, 622 11, 622 11, 622 20/20 1,019,863 3.86
G4 800 11, 646 11, 646 11, 646 20/20 1,350,718 5.20
G5 800 11, 631 11, 631 11, 631 20/20 390,206 1.56
G6 800 2178 2178 2178 20/20 383,868 1.32
G7 800 2006 2006 2006 20/20 1,395,985 6.54
G8 800 2005 2005 2005 20/20 1,875,952 8.71
G9 800 2054 2054 2054 20/20 951,896 3.98
G10 800 2000 2000 2000 20/20 4,485,966 16.54
G11 800 564 564 564 20/20 615,351 1.22
G12 800 556 556 556 20/20 2,139,865 4.03
G13 800 582 582 582 20/20 4,343,986 12.83
G14 800 3064 3064 3064 20/20 186,402,369 399.00
G15 800 3050 3050 3050 20/20 10,318,509 20.98
G16 800 3052 3052 3052 20/20 11,562,176 18.84
G17 800 3047 3047 3047 20/20 10,176,312 17.67
G18 800 992 992 992 20/20 5,113,729 8.36
G19 800 906 906 906 20/20 8,045,321 13.31
G20 800 941 941 941 20/20 1,665,698 3.08
G21 800 931 931 931 20/20 8,799,569 15.98
G22 2000 13, 359 13, 359 13, 359 20/20 33,354,509 77.62
G23 2000 13, 344 13, 344 13, 344 20/20 193,299,349 423.37
G24 2000 13, 337 13, 337 13, 337 20/20 117,515,927 355.75
G25 2000 13, 340 13, 340 13, 340 20/20 186,242,197 518.51
G26 2000 13, 328 13, 328 13, 328 20/20 207,731,604 566.09
G27 2000 3341 3341 3341 20/20 83,585,108 242.36
G28 2000 3298 3298 3298 20/20 155,389,382 417.02
G29 2000 3405 3405 3405 20/20 133,970,925 396.30
G30 2000 3413 3413 3413 20/20 168,679,922 397.76
G31 2000 3310 3310 3310 20/20 291,170,765 699.45
G32 2000 1410 1410 1410 20/20 23,254,862 81.50
G33 2000 1382 1382 1382 20/20 23,297,136 83.86
G34 2000 1384 1384 1384 20/20 15,079,978 51.16
G35 2000 7686 7687 7685.60 (0.73) 2/20 464,655,253 955.36
G36 2000 7680 7680 7677.50 (1.39) 3/20 523,019,635 1168.16
G37 2000 7691 7691 7688.05 (1.68) 3/20 430,916,535 943.63
G38 2000 7687 7688 7688 20/20 432,656,982 922.66
G39 2000 2408 2408 2408 20/20 319,652,386 681.56
G40 2000 2400 2400 2399.60 (1.20) 18/20 598,412,339 1189.28
G41 2000 2405 2405 2405 20/20 269,852,348 551.15
G42 2000 2481 2481 2478.45 (2.35) 8/20 704,523,691 1499.25
G43 1000 6660 6660 6659 20/20 2,503,251 2.54
G44 1000 6650 6650 6650 20/20 1,957,894 1.99
G45 1000 6654 6654 6654 20/20 5,189,575 4.23
G46 1000 6649 6649 6649 20/20 9,937,093 10.71
G47 1000 6657 6657 6657 20/20 3,121,725 3.39
G48 3000 6000 6000 6000 20/20 14,139 0.08
G49 3000 6000 6000 6000 20/20 17,254 0.14
G50 3000 5880 5880 5880 20/20 151,065 0.96
G51 1000 3848 3848 3848 20/20 73,478,520 246.21
G52 1000 3851 3851 3851 20/20 58,230,269 237.38
G53 1000 3850 3850 3849.55 (1.11) 17/20 135,679,730 513.96
G54 1000 3852 3852 3851.25 (1.37) 15/20 145,348,108 533.51
G55 5000 10, 299 10,299 10,291.75 (4.34) 3/20 226,942,156 2462.48
G56 5000 4016 4017 4008.60 (4.50) 2/20 182,963,792 1903.68
G57 5000 3492 3494 3488.70 (2.55) 2/20 398,023,655 4185.56
G58 5000 19, 263 19,276 19,266.00 (4.82) 1/20 408,756,239 4189.35
G59 5000 6078 6085 6070.45 (6.96) 1/20 412,389,566 4315.65
G60 7000 14, 176 14,186 14,173.50 (7.20) 1/20 346,689,235 5123.08
G61 7000 5789 5796 5776.00 (8.95) 2/20 302,598,635 4855.15
G62 7000 4868 4866 4860.20 (1.66) 1/20 312,598,632 4956.31
G63 7000 26, 997 27,018 26,993.60 (9.14) 2/20 468,562,389 5894.12
G64 7000 8735 8735 8717.95 (9.45) 1/20 456,894,123 5623.20
G65 8000 5558 5560 5555.40 (3.03) 2/20 281,568,423 5012.39
G66 9000 6360 6364 6353.70 (5.87) 1/20 298,562,318 6325.21
G67 10, 000 6940 6944 6937.30 (3.30) 2/20 198,562,387 6215.32
G70 10, 000 9541 9548 9539.60 (6.43) 1/20 201,259,836 6932.56
G72 10, 000 6998 6990 6979.70 (6.58) 1/20 212,589,563 7125.31
G77 14, 000 9926 9902 9890.80 (6.40) 1/20 186,523,694 7012.36
G81 20, 000 14, 030 14,010 13,993.20 (8.19) 1/20 102,369,845 6923.56
Q. Wu et al. / Applied Soft Computing 34 (2015) 827837 833

Table 3
Computational results of TSHEA on the cubic lattices instances.

Instance |V| fprev TSHEA

fbest favg () Hits Iter Time

sg3dl101000 1000 896 896 896 20/20 7,423,696 8.36


sg3dl102000 1000 900 900 900 20/20 2,789,563 3.15
sg3dl103000 1000 892 892 892 20/20 5,326,123 6.07
sg3dl104000 1000 898 898 898 20/20 7,986,325 9.10
sg3dl105000 1000 886 886 886 20/20 5,982,319 6.85
sg3dl106000 1000 888 888 888 20/20 3,956,238 4.02
sg3dl107000 1000 900 900 900 20/20 43,743,217 49.39
sg3dl108000 1000 882 882 882 20/20 7,516,569 8.10
sg3dl109000 1000 902 902 902 20/20 6,589,234 7.63
sg3dl1010000 1000 894 894 894 20/20 14,346,780 18.38
sg3dl141000 2744 2446 2446 2446 20/20 100,235,554 313.56
sg3dl142000 2744 2458 2458 2458 20/20 48,529,988 170.56
sg3dl143000 2744 2442 2442 2442 20/20 125,548,762 450.36
sg3dl144000 2744 2450 2450 2449.40 (0.91) 14/20 351,239,987 1388.34
sg3dl145000 2744 2446 2446 2446 20/20 50,214,223 163.58
sg3dl146000 2744 2452 2452 2451.40 (0.91) 14/20 594,829,576 1889.23
sg3dl147000 2744 2444 2444 2444 20/20 107,512,986 360.98
sg3dl148000 2744 2448 2448 2447.60 (0.80) 16/20 361,519,763 1381.39
sg3dl149000 2744 2426 2428 2426.30 (0.71) 3/20 526,984,125 1862.34
sg3dl1410000 2744 2458 2460 2458.40 (0.80) 4/20 312,536,932 1060.38

4. Analysis best solution value indicates that TS dominates TSN1 for all the
11 instances and performs better than TSN2 for 8 instances. The
In this section, we highlight the benets of the key components Wilcoxon rank-sum tests demonstrate that TS is able to produce
to the high performance of the proposed TSHEA algorithm. The signicantly better results than TSN1 and TSN2 for most of the
effects of the combination operator, neighborhood combination instances. These observations disclose the usefulness of the neigh-
and parameter settings are examined. In addition, the time-to- borhood combination in the tabu search algorithm.
target analysis is also used to evaluate the performance of different
algorithms. 4.3. Parameter tuning analysis

4.1. Solution combination operator analysis This experiment evaluates the inuence of parameter settings
to the performance of the TSHEA algorithm. To conduct this exper-
The operator for combining solutions plays a key role to the iment, we give different parameter settings to the tabu tenure tl,
performance of the proposed TSHEA algorithm. To verify its effec- the improvement cutoff and the tabu search depth . The average
tiveness, we compare TSHEA with its underlying tabu search solution values over 20 runs for each of the 11 used instances are
component (see Algorithm (3)). Tested on the chosen 11 hard collected.
instances, both algorithms are run under exactly the same time Fig. 1 depicts the box plots of the experimental results. The
limit as described in Section 3.2. Table 6 lists computational results box plots graphically depict the smallest result, lower quartile,
of TSHEA and TS on 11 hard instances, where better fbest between median, upper quartile, and the largest result obtained with each
TSHEA and TS are marked in bold. parameter value. The x-axis represents the parameter settings
From Table 6, we observe that TSHEA is able to reach better solu- and the y-axis represents for each instance the deviation of the
tions for 6 out of 11 instances than TS with slightly longer time. In average solution value from the best solution value, calculated as
addition, the observation that TSHEA gets higher hits than TS sug- (fbest favg )/fbest 100 %.
gests better robustness of TSHEA over TS. Finally, For each tested From Fig. 1, we nd that a careful choice of parameter settings is
problem instance, we apply the Wilcoxon rank-sum test with a indeed necessary for better algorithmic performance. For instance,
signicance factor of 0.05 to see whether there is any statistical dif- setting tl = 0.15|V| makes the average deviation from the best solu-
ference between the results achieved by our TSHEA and the results tion value fall within 0.1% for half of the instances, while the average
obtained by TS over 20 independent runs. The column p-value in deviation amounts to 0.22% when setting tl = 0.11|V|. In addition,
Table 6 shows the results reported by the Wilcoxon rank-sum test. we observe that the tabu search depth  seems more sensitive
From Table 6, we can observe that for 6 tested instances, TSHEA is than the tabu tenure tl and the tabu improvement cutoff . For
able to produce signicantly better results than TS, further revea- instance, setting  = 1e + 05 makes a quarter of instances with devi-
ling the merit of the solution combination operator included in our ations ranging from 0.2% to 0.75%, much larger than those obtained
TSHEA algorithm. by setting  = 1e + 06. Finally, for each specic parameter tl, or ,
if we take the median value of each parameter setting as a point
4.2. Tabu search analysis and line all these points, then we can nd the line achieves a U-
shape. This establishes the effectiveness of our chosen parameter
The neighborhood combination plays an important role in the settings.
tabu search algorithm. To assess its efcacy, we compare TS with its
variants that conduct search within a single neighborhood, either 4.4. Time-to-target analysis
using one-ip or exchange moves. The two variants are named as
TSN1 and TSN2 respectively. In order to further compare the efciency of the proposed TSHEA
Table 7 shows the results of TS, TSN1 and TSN2 on the 11 hard algorithm and other reference algorithms, we apply the time-to-
instances. All three algorithms were run under exactly the same target (TTT) analysis the empirical probability distribution for
time limit as described in Section 3.2. The rst observation of the the time to achieve a given target value [1]. Notice that a closed
834 Q. Wu et al. / Applied Soft Computing 34 (2015) 827837

Table 4
Computational comparisons between TSHEA, BLS, GLS and PR2 on the G-set instances.

Instance TSHEA BLS GES PR2

fbest favg fbest favg p-Value fbest favg p-Value fbest favg p-Value

G1 11,624 11,624.00 11,624 11,612.40 <0.001 11,624 11,624.00 1.000 11,624 11,624.00 1.000
G2 11,620 11,620.00 11,620 11,615.00 <0.001 11,620 11,620.00 1.000 11,620 11,620.00 1.000
G3 11,622 11,622.00 11,622 11,621.10 <0.001 11,622 11,622.00 1.000 11,620 11,620.00 1.000
G4 11,646 11,646.00 11,646 11,642.80 <0.001 11,646 11,646.00 1.000 11,646 11,646.00 1.000
G5 11,631 11,631.00 11,631 11,631.00 1.000 11,631 11,631.00 1.000 11,631 11,631.00 1.000
G6 2178 2178.00 2178 2178.00 1.000 2178 2178.00 1.000 2178 2178.00 1.000
G7 2006 2006.00 2006 2001.05 <0.001 2006 2006.00 1.000 2006 2006.00 1.000
G8 2005 2005.00 2005 2004.40 <0.001 2005 2005.00 1.000 2005 2005.00 1.000
G9 2054 2054.00 2054 2049.95 <0.001 2054 2054.00 1.000 2054 2054.00 1.000
G10 2000 2000.00 2000 1996.05 <0.001 2000 2000.00 1.000 2000 1999.80 <0.001
G11 564 564.00 564 564.00 1.000 564 564.00 1.000 564 564.00 1.000
G12 556 556.00 556 556.00 1.000 556 556.00 1.000 556 556.00 1.000
G13 582 582.00 582 582.00 1.000 582 582.00 1.000 582 582.00 1.000
G14 3064 3064.00 3064 3062.85 <0.001 3064 3064.00 1.000 3064 3062.60 <0.001
G15 3050 3050.00 3050 3050.00 1.000 3050 3050.00 1.000 3050 3049.30 <0.001
G16 3052 3052.00 3052 3051.10 <0.001 3052 3052.00 1.000 3052 3051.40 <0.001
G17 3047 3047.00 3047 3046.70 <0.001 3047 3047.00 1.000 3047 3046.40 <0.001
G18 992 992.00 992 991.70 <0.001 992 992.00 1.000 992 992.00 1.000
G19 906 906.00 906 904.55 <0.001 906 906.00 1.000 906 906.00 1.000
G20 941 941.00 941 941.00 1.000 941 941.00 1.000 941 941.00 1.000
G21 931 931.00 931 930.20 <0.001 931 931.00 1.000 931 931.00 1.000
G22 13,359 13,359.00 13,359 13,344.45 <0.001 13,359 13,359.00 1.000 13,359 13,354.50 <0.001
G23 13,344 13,344.00 13,344 13,340.60 <0.001 13,342 13,342.00 <0.001 13,342 13,331.60 <0.001
G24 13,337 13,337.00 13,337 13,329.80 <0.001 13,337 13,337.00 1.000 13,333 13,325.30 <0.001
G25 13,340 13,340.00 13,340 13,333.40 <0.001 13,340 13,340.00 1.000 13,339 13,328.20 <0.001
G26 13,328 13,328.00 13,328 13,320.00 <0.001 13,328 13,327.90 <0.001 13,326 13,312.30 <0.001
G27 3341 3341.00 3341 3332.25 <0.001 3341 3341.00 1.000 3336 3326.90 <0.001
G28 3298 3298.00 3298 3293.85 <0.001 3298 3298.00 1.000 3296 3288.90 <0.001
G29 3405 3405.00 3405 3388.20 <0.001 3405 3405.00 1.000 3405 3391.90 <0.001
G30 3413 3413.00 3412 3404.85 <0.001 3413 3413.00 1.000 3411 3404.80 <0.001
G31 3310 3310.00 3309 3305.30 <0.001 3310 3310.00 1.000 3306 3299.50 <0.001
G32 1410 1410.00 1410 1409.30 <0.001 1410 1410.00 1.000 1410 1404.60 <0.001
G33 1382 1382.00 1382 1380.10 <0.001 1382 1382.00 1.000 1382 1376.10 <0.001
G34 1384 1384.00 1384 1384.00 1.000 1384 1384.00 1.000 1384 1378.20 <0.001
G35 7687 7685.60 7684 7680.85 <0.001 7686 7685.55 0.775 7679 7670.80 <0.001
G36 7680 7677.50 7678 7673.60 <0.001 7680 7676.65 0.016 7671 7658.70 <0.001
G37 7691 7688.05 7689 7685.85 <0.001 7691 7690.10 <0.001 7682 7667.90 <0.001
G38 7688 7688.00 7687 7684.95 <0.001 7687 7687.00 <0.001 7682 7670.40 <0.001
G39 2408 2408.00 2408 2405.35 <0.001 2408 2408.00 1.000 2407 2391.10 <0.001
G40 2400 2399.60 2400 2394.60 <0.001 2400 2399.55 0.858 2399 2383.30 <0.001
G41 2405 2405.00 2405 2403.00 <0.001 2405 2405.00 1.000 2404 2388.90 <0.001
G42 2481 2478.45 2481 2475.40 <0.001 2481 2481.00 <0.001 2478 2466.20 <0.001
G43 6660 6660.00 6660 6658.15 <0.001 6660 6660.00 1.000 6660 6659.90 <0.001
G44 6650 6650.00 6650 6650.00 1.000 6650 6647.70 <0.001 6650 6649.90 <0.001
G45 6654 6654.00 6654 6652.15 <0.001 6654 6654.00 1.000 6654 6653.90 <0.001
G46 6649 6649.00 6649 6647.75 <0.001 6649 6649.00 1.000 6649 6648.80 <0.001
G47 6657 6657.00 6657 6654.35 <0.001 6657 6657.00 1.000 6657 6656.80 <0.001
G48 6000 6000.00 6000 6000.00 1.000 6000 6000.00 1.000 6000 6000.00 1.000
G49 6000 6000.00 6000 6000.00 1.000 6000 6000.00 1.000 6000 6000.00 1.000
G50 5880 5880.00 5880 5879.90 <0.001 5880 5880.00 1.000 5880 5880.00 1.000
G51 3848 3848.00 3848 3847.85 <0.001 3848 3848.00 1.000 3848 3846.40 <0.001
G52 3851 3851.00 3851 3850.85 <0.001 3851 3851.00 1.000 3851 3848.40 <0.001
G53 3850 3849.55 3850 3849.50 0.847 3850 3849.90 0.188 3850 3847.70 <0.001
G54 3852 3851.25 3852 3850.60 0.047 3852 3852.00 0.024 3851 3847.80 <0.001
G55 10,299 10,291.75 10,294 10,282.40 <0.001 10,265 10,234.00 <0.001
G56 4017 4008.60 4012 3998.65 <0.001 3981 3959.20 <0.001
G57 3494 3488.70 3492 3488.60 0.028 3472 3462.00 <0.001
G58 19,276 19,266.00 19,263 19,255.90 <0.001 19,205 19,182.00 <0.001
G59 6085 6070.45 6078 6067.90 0.126 6027 6006.20 <0.001
G60 14,186 14,173.50 14,176 14,166.80 <0.001 14,112 14,091.80 <0.001
G61 5796 5776.00 5789 5773.35 0.212 5730 5695.70 <0.001
G62 4866 4860.20 4868 4863.80 <0.001 4836 4830.20 <0.001
G63 27,018 26,993.60 26,997 26,980.70 <0.001 26,916 26,879.30 <0.001
G64 8735 8717.95 8735 8735.00 <0.001 8641 8594.10 <0.001
G65 5560 5555.40 5558 5551.20 <0.001 5526 5515.90 <0.001
G66 6364 6353.70 6360 6350.20 0.018 6314 6302.40 <0.001
G67 6944 6937.30 6940 6935.30 0.016 6902 6884.60 <0.001
G70 9548 9539.60 9541 9527.10 <0.001 9463 9434.00 <0.001
G72 6990 6979.70 6998 6935.30 <0.001 6946 6933.80 <0.001
G77 9902 9890.80 9926 9916.10 <0.001
G81 14,010 13,993.20 14,030 14,021.70 <0.001
Q. Wu et al. / Applied Soft Computing 34 (2015) 827837 835

Table 5
Computational comparisons between TSHEA, GES and PR2 on the cubic lattices instances.

Instance TSHEA GES PR2

fbest favg fbest favg p-Value fbest favg p-Value

sg3dl101000 896 896.00 896 896.00 1.000 896 894.60 <0.001


sg3dl102000 900 900.00 900 900.00 1.000 900 900.00 1.000
sg3dl103000 892 892.00 892 892.00 1.000 892 891.30 <0.001
sg3dl104000 898 898.00 898 898.00 1.000 898 898.00 1.000
sg3dl105000 886 886.00 886 886.00 1.000 886 885.40 <0.001
sg3dl106000 888 888.00 888 888.00 1.000 888 888.00 1.000
sg3dl107000 900 900.00 900 900.00 1.000 900 898.20 <0.001
sg3dl108000 882 882.00 882 882.00 1.000 882 881.20 <0.001
sg3dl109000 902 902.00 902 902.00 1.000 902 901.50 <0.001
sg3dl1010000 894 894.00 894 894.00 1.000 894 893.70 <0.001
sg3dl141000 2446 2446.00 2446 2445.90 <0.001 2444 2437.60 <0.001
sg3dl142000 2458 2458.00 2458 2458.00 1.000 2456 2452.40 <0.001
sg3dl143000 2442 2442.00 2442 2442.00 1.000 2438 2435.50 <0.001
sg3dl144000 2450 2449.40 2450 2449.50 0.639 2448 2440.00 <0.001
sg3dl145000 2446 2446.00 2446 2446.00 1.000 2444 2438.70 <0.001
sg3dl146000 2452 2451.40 2452 2451.50 0.639 2448 2442.30 <0.001
sg3dl147000 2444 2444.00 2444 2444.00 1.000 2440 2435.00 <0.001
sg3dl148000 2448 2447.60 2448 2447.50 0.592 2444 2438.90 <0.001
sg3dl149000 2428 2426.30 2426 2426.00 0.083 2422 2417.30 <0.001
sg3dl1410000 2460 2458.40 2458 2457.90 0.013 2454 2448.80 <0.001

Table 6
Comparative results of TSHEA and TS on 11 hard G-set instances.

Instance |V| TSHEA TS

fbest favg Hits Time fbest favg Hits Time p-Value

G32 2000 1410 1410 20/20 81.50 1410 1410 20/20 79.25 1.000
G33 2000 1382 1382 20/20 83.86 1382 1382 20/20 81.21 1.000
G35 2000 7687 7685.65 2/20 955.36 7683 7679.90 1/20 1536.23 <0.001
G40 2000 2400 2399.60 18/20 1189.28 2400 2397.50 2/20 815.32 <0.001
G41 2000 2405 2405 20/20 551.15 2405 2404.10 14/20 826.71 0.010
G42 2000 2481 2478.45 8/20 1499.25 2481 2477.30 1/20 1238.26 0.129
G55 5000 10,299 10,290.75 3/20 2462.48 10,295 10,290.90 1/20 2895.32 0.724
G57 5000 3494 3488.70 2/20 4185.56 3490 3488.30 4/20 4253.21 0.044
G62 7000 4866 4860.20 1/20 4956.31 4862 4859.50 1/20 5269.12 0.594
G65 8000 5560 5555.40 2/20 5012.39 5546 5542.50 4/20 4852.10 <0.001
G66 9000 6364 6353.70 1/20 6325.21 6340 6334.80 2/20 4218.96 <0.001

form index [21] could be another tool to compare two or more From Fig. 2, we observe that the blue lines of TSHEA are strictly
algorithms. Since the source codes of the BLS and GES algorithms above the green line of PR2 for both instances G19 and G25, indicat-
are not available, we conduct this TTT test on TSHEA and PR2 algo- ing that either TSHEA always has a higher probability than PR2 to
rithms. reach the target value within a given time period or, equivalently,
We carry out the TTT experiment by performing 200 indepen- that TSHEA uses shorter time than PR2 to acquire the same proba-
dent runs for TSHEA and PR2 on each instance. The time needed bility of successfully reaching the target value. Specically, for the
to attain an objective value at least as good as a given target value instance G19, TSHEA has a probability of 70% to reach the given
for each run is recorded. Then we sort the recorded times in an target value within roughly 70 s while PR2 takes 180 s to acquire
increasing order so that ti represents the ith lower time and a the same performance. For the instance G25, TSHEA takes less than
probability pi = (i 0.5)/200 is associated to each time ti . Finally, 70 s to reach the given target value for each independent run, while
the points (ti , pi ) are plotted. Fig. 2 shows the results of the TTT within this time limit PR2 can only have a probability of less than
experiment for TSHEA and PR2 on the instances G19 (left) and G25 40% to successfully reach the target value. The TTT experiments are
(right). also conducted on other instances and the similar observation can

Table 7
Comparative results of the combined neighborhood and two single neighborhoods on 11 G-set instances.

Instance |V| TS TSN1 TSN2

fbest favg Hits Time fbest favg Hits Time p-Value fbest favg Hits Time p-Value
G32 2000 1410 1410 20/20 79.95 1398 1393.50 1/20 512.36 <0.001 1410 1410 20/20 485.32 1.000
G33 2000 1382 1382 20/20 81.21 1372 1366.40 1/20 1612.25 <0.001 1382 1381.90 19/20 689.15 0.342
G35 2000 7683 7679.90 1/20 1536.23 7679 7674.55 1/20 781.23 <0.001 7683 7676.20 1/20 1203.95 0.04
G40 2000 2400 2397.50 2/20 815.32 2396 2389.35 1/20 286.31 <0.001 2396 2391.20 1/20 918.21 <0.001
G41 2000 2405 2404.10 14/20 826.71 2404 2394.50 3/20 918.32 <0.001 2404 2398.70 4/20 886.38 <0.001
G42 2000 2481 2477.30 1/20 1238.26 2477 2471.65 1/20 1236.21 <0.001 2477 2470.45 1/20 1425.68 <0.001
G55 5000 10,295 10,290.90 1/20 2895.32 10,289 10,281.35 1/20 3612.35 <0.001 10,288 10,281.30 1/20 6231.08 <0.001
G57 5000 3490 3488.30 4/20 4253.21 3432 3426.60 5/20 3598.12 <0.001 3484 3477.80 3/20 3892.25 <0.001
G62 7000 4862 4859.50 1/20 5269.12 4786 4779.80 2/20 4716.32 <0.001 4856 4838.80 1/20 4952.13 <0.001
G65 8000 5546 5542.50 4/20 4852.10 5466 5450.20 1/20 2896.32 <0.001 5520 5498.00 1/20 1602.31 <0.001
G66 9000 6340 6334.80 2/20 4218.96 6248 6240.00 1/20 3893.21 <0.001 6306 6280.50 1/20 3212.98 <0.001
836 Q. Wu et al. / Applied Soft Computing 34 (2015) 827837

Fig. 1. Box and whisker plot of the results obtained with different settings for each sensitive parameter.

1 1
TSHEA TSHEA
PR2 PR2
0.8 0.8
Probability

Probability

0.6 0.6

0.4 0.4

Instances: g19 Instances: g25


0.2 Target: 906 0.2 Target: 13316

0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Time to target value Time to target value
(a) Instance: G19 (b) Instance: G25

Fig. 2. Empirical probability distribution for the time to achieve a target value.

be made. Therefore, this experiment shows that TSHEA dominates exchange moves. The solution combination operator inherits com-
PR2 with respect to computational efciency. mon solution components from parent solutions and completes the
remaining part using a quality-and-distance evaluation, which can
5. Conclusion be considered as an innovative variant of the traditional uniform
crossover.
In this paper, we propose a tabu search based hybrid evolution- Extensive experiments on a total of 91 instances demonstrate
ary algorithm for solving the NP-hard max-cut problem. The tabu that TSHEA is capable of improving the best known results for
search component alternates between two search phases, one of 15 instances and matching them for all except 4 instances, while
which is based on one-ip move and the other relies on constrained attaining this performance within comparable computing efforts.
Q. Wu et al. / Applied Soft Computing 34 (2015) 827837 837

Statistical comparisons indicate that the proposed TSHEA algo- [10] F. Glover, M. Laguna, Tabu Search, Kluwer Academic Publishers, Boston, 1997.
rithm is signicantly better than several currently state-of-the-art [11] F. Glover, J.K. Hao, Efcient evaluation for solving 01 unconstrained quadratic
optimization problems, Int. J. Metaheuristics 1 (1) (2010) 310.
algorithms. [12] J.K. Hao, Memetic algorithms in discrete optimization, in: F. Neri, C. Cotta, P.
In addition, we examine several important components of Moscato (Eds.), Handbook of Memetic Algorithms. Studies in Computational
TSHEA, including the solution combination operator, the combined Intelligence, vol. 379, 2012, pp. 7394 (Chapter 6).
[13] J. Kneis, P. Rossmanith, A new satisablity algorithm with applications to max-
neighborhood and the parameter settings. The analysis reveals cut. Technical Report AIB-2005-08, Department of Computer Science, RWTH
their merit to the high performance of the TSHEA algorithm. Aachen, 2005.
[14] G. Kochenberger, J.K. Hao, Z. L, H. Wang, F. Glover, Solving large scale max cut
problems via tabu search, J. Heuristics 19 (4) (2013) 565571.
Acknowledgements [15] K. Krishnan, J. Mitchell, A semidenite programming based polyhedral cut and
price approach for the maxcut problem, Comput. Optim. Appl. 33 (1) (2006)
We are grateful to the reviewers whose comments have helped 5571.
[16] R. Marti, A. Duarte, M. Laguna, Advanced scatter search for the max-cut prob-
to improve the quality of the paper. This work was supported
lem, INFORMS J. Comput. 21 (1) (2009) 2638.
by National Natural Science Foundation of China (Grant No. [17] N. Mladenovic, P. Hansen, Variable neighborhood search, Comput. Oper. Res.
71131004, 71332001, 71401059, 61370183), Program for New Cen- 24 (11) (1997) 1097C1100.
[18] F. Neri, C. Cotta, P. Moscato (Eds.), Handbook of Memetic Algorithms. Studies
tury Excellent Talents in University (NCET 2013) and Northwestern
in Computational Intelligence, vol. 379, Springer, 2011.
Polytechnical University (Grant No. 15GH0312, 3102015RW007, [19] G. Palubeckis, Application of multistart tabu search to the MaxCut problem,
W016233). Inform. Technol. Control 2 (31) (2004) 2935.
[20] D.C. Porumbel, J.K. Hao, P. Kuntz, An evolutionary approach with diversity guar-
antee and well-informed grouping recombination for graph coloring, Comput.
References Oper. Res. 37 (10) (2010) 18221832.
[21] C.C. Ribeiro, I. Rosseti, R. Vallejos, Exploiting run time distributions to compare
[1] R.M. Aiex, M.G.C. Resende, C.C. Ribeiro, TTT plots: a perl program to create sequential and parallel stochastic local search algorithms, J. Global Optim. 54
time-to-target plots, Optim. Lett. 1 (2007) 355366. (2012) 405429.
[2] E. Arriz, O. Olivo, Competitive simulated annealing and tabu search algo- [22] A.D. Scott, G.B. Sorkin, Solving sparse semi-random instances of Max-Cut and
rithms for the max-cut problem, in: GECCO 2009 Proceedings of the 11th Max-CSP in linear expected time, Research Report 23417 (W0411-056), IBM
Annual Conference on Genetic and Evolutionary Computation, 2009, pp. 1797 Research division, Thomas J. Watson Research Center (2004).
1798. [23] V.P. Shylo, O.V. Shylo, Solving the maxcut problem by the global equilibrium
[3] F. Barahona, M. Grtschel, M. Jnger, G. Reinelt, An application of combinatorial search, Cybern. Syst. Anal. 46 (5) (2010) 744754.
optimization to statistical physics and circuit layout design, Oper. Res. 36 (1988) [24] V.P. Shylo, O.V. Shylo, B.V.A. Roschyn, Solving weighted max-cut problem by
493513. global equilibrium search, Cybern. Syst. Anal. 48 (4) (2012) 563567.
[4] U. Benlic, J.K. Hao, A multilevel memetic approach for improving graph k- [25] T. Sttzle, Iterated local search for the quadratic assignment problem, Eur. J.
partitions, IEEE Trans. Evol. Comput. 15 (5) (2011) 624642. Oper. Res. 174 (2006) 15191539.
[5] U. Benlic, J.K. Hao, Breakout local search for the max-cut problem, Eng. Appl. [26] P. Venkatesh, A. Singh, Two metaheuristic approaches for the multiple traveling
Artif. Intell. 26 (3) (2013) 11621173. salesperson problem, Appl. Soft Comput. 26 (2015) 7489.
[6] S. Burer, R.D.C. Monteiro, Y. Zhang, Rank-two relaxation heuristics for max- [27] Y. Wang, Z. Lu, J.K. Hao, F. Glover, Path relinking for unconstrained
cut and other binary quadratic programs, SIAM J. Optim. 12 (2) (2002) binary quadratic programming, Eur. J. Oper. Res. 223 (3) (2012)
503521. 595604.
[7] F.D. Crocea, M.J. Kaminskib, V.Th. Paschosc, An exact algorithm for MAX-CUT [28] Q. Wu, J.K. Hao, A memetic approach for the max-cut problem., in: C. Coello,
in sparse graphs, Oper. Res. Lett. 35 (3) (2007) 403408. V. Cutello, K. Deb, S. Forrest, G. Nicosia, M. Pavone (Eds.), PPSN 2012, Lecture
[8] P. Festa, P.M. Pardalos, M.G.C. Resende, C.C. Ribeiro, Randomized heuristics for Notes in Computer Science, vol. 7492, 2012, pp. 297306.
the max-cut problem, Optim. Methods Softw. 7 (2002) 10331058. [29] Q. Wu, J.K. Hao, Memetic search for the max-bisection problem, Comput. Oper.
[9] M.X. Goemans, D.P. Williamson, Improved approximation algorithms for max- Res. 40 (1) (2013) 166179.
cut and satisability problems using semidenite programming, J. ACM 42 (6) [30] Q. Wu, J.K. Hao, A review on algorithms for maximum clique problems, Eur. J.
(1995) 11151145. Oper. Res. 242 (2015) 693709.

You might also like